Simultaneous localization and mapping (SLAM) must remain reliable when sensing suites and operating conditions vary across platforms and deployments. Beyond correspondence degradation, a dominant deployment failure mode is misweighted constraints: under distribution shift, uncertainty estimates can become miscalibrated, allowing a small set of overconfident factors to dominate iterative optimization and destabilize inference. This article presents conformal-calibrated foundation-factor graph SLAM ( ${\mathrm{C}\mathrm{F}}^2\text{-}SLAM$), a sensor-agnostic framework that combines frozen foundation representations with lightweight probabilistic factor heads that emit explicit residuals and covariances, and a classical factor-graph back-end for principled multi-modal fusion. To mitigate systematic misweighting under shift, an online conformal calibration layer is introduced to rescale factor covariances by aligning empirical residual quantiles with target quantiles on a per-factor-family basis. Loop closure is further integrated through foundation-descriptor retrieval for candidate proposal and conservative geometric verification for graph insertion, controlling false loop constraints without relying on dataset-specific place-recognition supervision. Across heterogeneous benchmarks spanning monocular, stereo, red-green-blue-depth (RGB-D), and visual-inertial settings, ${\mathrm{C}\mathrm{F}}^2\text{-}SLAM$ operates without retraining and shows improved robustness trends under zero-shot transfer, consistent with stabilized factor weighting.